Vibe Coding

Vibe coding is the practice of building software by describing intent in natural language and letting AI tools generate the code. Common tools include Cursor, Claude Code, Devin, and Replit Agent. The term was coined by Andrej Karpathy in a February 2025 tweet: "There's a new kind of coding I call 'vibe coding,' where you fully give in to the vibes, embrace exponentials, and forget that the code even exists." It captures the shift from typing code character-by-character to describing outcomes and reviewing AI-generated implementations. It's the most visible name for the broader shift in how software gets built in the AI era.

The Karpathy origin (February 2025):

Andrej Karpathy (former Tesla AI lead, OpenAI co-founder, prolific AI educator) described vibe coding as a new mode of software development where the developer:

• Describes the desired outcome to AI in natural language.
• Reviews and accepts/rejects AI-generated code.
• Iterates by describing changes rather than editing manually.
• "Forgets the code exists", focuses on outcomes, not implementation.

The post went viral, "vibe coding" became the cultural term for AI-driven software development through 2025-2026, with Cursor, Claude Code, and Replit Agent as primary tools.

The tools that enable vibe coding:

Cursor: AI-native code editor; describes intent, AI generates and modifies code across files.

Claude Code: command-line AI coding agent.

GitHub Copilot Workspace: pull-request-level AI generation.

Devin (Cognition): autonomous AI software engineer.

Replit Agent: build apps from natural language description.

v0 (Vercel): generate React components and apps from prompts.

Lovable / Bolt / others: AI app builders.

Each has different opinions on how much vibe vs how much code review.

What vibe coding works well for:

Prototyping: get a working version fast to validate ideas.

Greenfield projects: less existing context = better AI performance.

Common patterns: AI knows React + TypeScript + standard backend patterns well.

Web apps and consumer interfaces: well-trodden territory for AI tools.

Scripts and automation: simple, well-defined tasks.

Learning new technologies: AI can scaffold and explain.

What vibe coding struggles with:

Complex existing codebases: AI loses context across large codebases.

Performance-critical systems: optimization requires deep understanding.

Novel architectures: AI follows known patterns; novel work needs human design.

Security-sensitive code: AI generates security bugs at non-trivial rates.

Debugging complex issues: AI is often weaker at root-cause analysis than implementation.

Production scale: AI-generated code sometimes works in dev, breaks at scale.

The vibe coding skill set:

The new developer skills emerging:

Prompt engineering for code: writing prompts that produce good code.

Architecture judgment: knowing what to ask AI to build vs build yourself.

Code review at scale: reviewing AI-generated code rapidly.

Test-driven AI development: writing tests that constrain AI implementations.

Debugging AI mistakes: identifying where AI went wrong.

Tool selection: knowing which AI tool fits which task.

Specification skills: describing requirements clearly enough for AI execution.

The productivity gains (and exaggerations):

Real gains:

• 10-30% productivity improvement for senior devs using Cursor / Copilot (industry studies).
• 50-100%+ for some tasks (prototyping, boilerplate).
• 2-5x productivity for less experienced developers (more upside from AI assistance).

Exaggerated claims:

• "10x productivity", true for some tasks, not all.
• "Replaces engineers", junior tasks shift; senior judgment still needed.
• "Anyone can be a developer now", partially true; building production software still requires expertise.

The startup implication:

Founder coding velocity: founders without engineering background can build more than before. "Solo founders shipping enterprise products" is increasingly viable.

Engineering team size: small teams can ship more code; some companies running with 2-5 engineers what would have required 20-50.

Code quality concerns: shipped fast doesn't mean shipped right. Production quality requires care.

Hiring shifts: "AI-native engineers" who use vibe coding effectively are higher-leverage; engineers who don't are less competitive.

Open-source acceleration: open-source projects move faster with AI contributions.

The cultural debate:

Skeptics: "vibe coding" produces shallow understanding, brittle code, AI dependency.

Enthusiasts: vibe coding is a paradigm shift on par with going from assembly to high-level languages.

Pragmatists: vibe coding is a useful tool for some tasks; traditional coding still essential for others.

Reality: all three are partially right. The truth is nuanced and depends on context.

What founders get wrong: Two opposite mistakes: (1) dismissing vibe coding as overhyped (it's real productivity gain), or (2) believing it eliminates the need for engineering judgment (it doesn't). The right discipline: use AI tools effectively; maintain engineering judgment for what matters; ship faster but ship right.

Related: AI Agent · Large Language Model · Generative AI · Prompt Engineering

FAQ

What is vibe coding?
The practice of building software by describing intent in natural language and letting AI tools (Cursor, Claude Code, Devin, Replit Agent) generate the code. Coined by Andrej Karpathy in February 2025.

What tools are used for vibe coding?
Cursor (AI-native editor), Claude Code (command-line agent), GitHub Copilot Workspace, Devin (Cognition), Replit Agent, v0 (Vercel), Lovable, Bolt. Different tools have different vibe-vs-review balance.

Does vibe coding produce production-quality code?
For some tasks yes, others no. Works well for prototyping, common patterns, web apps, scripts. Struggles with complex existing codebases, performance-critical systems, novel architectures, security-sensitive code, complex debugging. Code review still essential.

Will vibe coding replace software engineers?
No, but it shifts what engineers do. Productivity gains: 10-30% for senior devs; up to 2-5x for less experienced. Engineering judgment for architecture, security, performance, novel work still essential. "AI-native engineers" who use vibe coding effectively are higher-leverage; engineers who don't are less competitive.